On 07/22/2011 11:29 PM, David Chapman wrote:
On 7/22/2011 8:38 AM, Andy Canfield wrote:
Had what seems to be a bright idea. It is a bright idea for a
Subversion server on Linux or OS X; AFAIK this idea has no relevance
to Windows:
Take note of the user and group that Apache runs as. Call this
combination APACHE, meaning APACHE_USER and APACHE_GROUP.
Whatever mod_dav_svn does to any repository will be done by APACHE.
Whatever WebSVN does to any repository will be done by APACHE.
If I set the svnserve program to be owned by APACHE, and setuid and
setgid, then whatever svnserve does to any repository will also be
done by APACHE. Only root, or the APACHE user, can make this change
to the svnserve program binary.
If I set the svnadmin program to be owned by APACHE, and setuid and
setgid, then whatever svnadmin does to any repository will also be
done by APACHE.
If I set the svnlook program to be owned by APACHE, and setuid and
setgid, then whatever svnlook does to any repository will be done by
APACHE.
So the three access paths - http:, svn:, and direct - will all
operate using the same user and group.
You don't want to do this to any program, such as 'svn', which relies
on user authentication, since it needs to know the actual user that
is running the program. But any program which operates directly on
the repository can be set this way and that ensures that the
repository is always manipulated, at a low level, by the same user
all the time.
So a post-installation setup would include:
* sudo bash
cd /usr/bin
chown APACHE_USER svnadmin svnlook svnserve
chgrp APACHE_GROUP svnadmin svnlook svnserve*
This idea also allows me to make the the respository itself
accessable only by APACHE. This ensures that the repository can only
be manipulated by Subversion code itself (or by root):
* sudo bash
mkdir /var/svn
chown APACHE_USER**/var/svn*
* chgrp APACHE_GROUP**/var/svn*
* chmod 0700 /var/svn*
If I also do this:
* sudo bash
rm /usr/bin/svnadmin*
then all repository creation would have to be done via something like
WebSVN, which I assume requires authentication.
How does that sound to you guys? Where are the flaws in this idea?
Thank you very much.
It prevents privately owned and managed repositories. Normally,
individual users can set up their own repositories for their own
(personal) projects or files. If you are planning to do this on a
server that you wish to lock down, then I see no reason why it would
cause problems, but if users can login and do other work on this
machine, you are constraining them.
Ah, I see a lack of terminology here. Consider the respository located
at /var/svn/sample, the repository named "sample" and accessed through
the browser as "http://example.com/svn/sample";. What is the name for
"/var/svn"? Is it a "Repository Collection"? The config entry is
"SVNParentPath", so let's call it the "SVNParent".
If you have an account on that server computer, and you don't have root
access, and you want your own pesonal repository collection, then even
today you can't use http: access to it, either because you can't
configure Apache or because Apache is already configured for the
standard SVNParent.
The suggestion would make the SVNParent accessable only through the
standard software, never directly.
Any user could create his own respository, but it would be in the
standard SVNParent.
If /usr/bin/svnserve is owned by www-data (APACHE) and setuid, then you
can't run /usr/bin/svnserve to access your own personal SVNParent. It
can access your own repository in the standard SVNParent, but not your
own SVNParent.
However, it appears that it can be done using two Linux commands:
* copy /usr/bin/svnserve ~/bin/svnserve
copy /usr/bin/svnadmin ~/bin/svnadmin*
The copies are owned by you and are no longer setuid. So you can run
~/bin/svnserve and it will operate under your user name and manage your
own personal SVNParent. (I have not tested this yet, but it "should work").
On Fri, 22 Jul 2011 13:24:55 -0400 Nico Kadel-Garcia wrote:
Sounds dangerous and untested. svnserve, for example, is not designed
to be run suid. It's not necessarily handling UID versus EUID
correctly for this. (It might, it's not tested.)
That sounds like a serious concern; no way to be sure. But for me it
would be worth the experiment.
